The new Jotunheim: OpenAI's Stargate and xAI's Colossus

Tech's next realm of giants will be built more overtly on atoms, not bits

OpenAI just revealed a new venture called ‘Stargate’ in partnership with SoftBank that will direct $500 billion over the course of the next four years to build out the data centre infrastructure to power its technology in the US, beginning with a facility in Texas that has already broken ground.

Stargate’s active partners are primarily OpenAI (operations) and SoftBank (finances), but it’s a collaborative effort that includes equity investment from Oracle and MGX as well, and technology support from Arm, Microsoft, and NVIDIA, alongside Oracle and OpenAI.

This isn’t a surprise move: Reports from the past couple of years had OpenAI’s Sam Altman traipsing around the world looking for capital partners to help fund infrastructure build-out to the tune of trillions, not just hundreds of billions. And the stakes continue to grow, with rivals like Elon Musk out there rough-hewing multi-billion dollar data centers virtually out of thin air in the outskirts of Memphis.

Musk’s Colossus supercomputer is massive in scope, and the billionaire technologist’s ambitions are even bigger: He’s said that it will eventually expand to house over 1 million GPUs, 10x the 100,000 that call it home today. Even so, the $500 billion figure that OpenAI dropped yesterday dwarfs the financial investment Musk and co. have sunk into xAI’s infrastructure to date, as well as its projected plans.

Already, Musk is attempting to throw cold water on the apparent disparity in capital expenditure between Stargate and his own ambitions. He posted on X on Tuesday following the announcement that the new consortium doesn’t “actually have the money,” with “well under $10B secured” by finance lead SoftBank as of the announcement according to the Tesla CEO.

That’s a bit rich coming from the same guy who once infamously tweeted ‘funding secured’ when actually that was super not the case. Still, he has a point here that the eye-watering number is at this point mostly aspirational rather than concrete.

Certainly, challenges remain in terms of getting to these lofty investment goals, but whether it’s $10, 20, 50 or 500 billion, the fact is we’re going to see increasing amounts of money shovelled in to these kinds of infrastructure projects.

There are reasons to doubt the sustainability of an AI-supporting infrastructure boom. I think there’s an argument to be made that we’re in the midst of an AI bubble that may pop to some degree, and people in the industry I respect to a great degree are also saying that there are diminishing returns in terms of applying compute and spend at the training part of the AI tech stack. Just this week we saw DeepSeek AI roll out a relatively lightweight, open source model that claims to be able to outperform some of OpenAI’s best and biggest, so it also seems increasingly like the foundational model game is going to be rapidly commoditized.

On top of that, unlike even the highly-commoditized cloud providers, AI models are actually almost trivially easy and cheap to swap in and out. With marginal switching costs and a demonstrated ability to leapfrog larger models with small ones in terms of performance in short order, what’s the value in massive underlying infrastructure spend?

I think the answer lies in inference time compute, and how that will increasingly be a valuable differentiator – particularly when it comes to running multiple inference tasks targeting a single result, and then using that array of results to achieve an optimal outcome through orchestration, combination and revision.

A lot of product layer companies are doing this kind of thing already, and delivering real value there. I think we’ll increasingly see foundation model providers like OpenAI explore what else they can do at this layer of the stack – and especially as they work out what kind of defensible advantage they can bring to the table by amassing an unmatchable pool of sheer hardware horsepower to drive those inference-time gains.